← X

@gopikl @codyplof Sometimes I take for granted how quickly we can ship great pro

X · bcherny · April 4, 2026
A discussion emerged around Anthropic's approach to product development, noting the contrast between rapid feature shipping and the technical complexity of maintaining high-throughput inference infrastructure at scale. Multiple users criticized Claude's subscription plans for restrictive usage limits, unclear refund policies, and limitations on third-party tool integration, with some switching to competing models due to cost and functionality concerns.

Detailed Analysis

Anthropic's product shipping velocity has emerged as a defining characteristic of the company's competitive posture in early 2026, with internal data showing 74 features released across 52 days between February and March, and over 120 features in 90 days. A tweet from Boris Cherny, an Anthropic engineer, directed at colleagues acknowledges this pace while candidly admitting the asymmetry between shipping product and sustaining the underlying infrastructure: building a high-throughput inference and API stack at scale is a substantially harder engineering challenge. The comment reflects genuine internal tension — teams can iterate on features quickly, but the physical and architectural demands of serving those features to a massive, rapidly growing user base introduce constraints that product velocity alone cannot resolve.

The thread that surrounds Cherny's tweet captures a user community in visible turmoil, largely triggered by Anthropic's decision to restrict third-party API access through consumer subscription plans — a move that effectively shut down tools like OpenClaw that had routed traffic through consumer accounts rather than the paid API. Users express frustration across several dimensions: rate limits on the $20/month plan described as impractically low (roughly 8 messages before a five-hour cooldown), difficulty obtaining refunds for unused subscriptions, broken credit promises, and a general sense that Anthropic's policy changes prioritized enterprise revenue over the developer community that built tooling around its models. The complaints are significant not merely as customer service noise but as a signal that Anthropic's pricing and access architecture has created friction with exactly the power users and builders who generate organic adoption.

Anthropic's infrastructure scaling challenge, which Cherny alludes to, is directly connected to the product releases that preceded this public moment. Claude Opus 4.6, launched in early February 2026 with a one-million-token context window, and Claude Sonnet 4.6, which followed two weeks later with 30–50% speed improvements, represent genuinely demanding workloads at scale. Each new capability — longer context, computer control via Cowork, phone access through Claude Dispatch — multiplies per-request compute requirements. The decision to restrict subscription-tier API access is therefore not purely commercial; it reflects a real infrastructure constraint that makes it economically and technically unsustainable to allow unlimited agentic or automated workloads on flat-fee consumer plans, regardless of how those workloads are routed.

The broader competitive context makes Anthropic's execution challenge more acute. Commenters in the thread point to OpenAI's Codex, Google's Gemma, Alibaba's Qwen, and MiniMax as alternatives they are actively migrating to, and at least one user reports that a Claude subscription has been formally banned from an internal enterprise project. This competitive pressure matters because Anthropic's shipping velocity — while objectively impressive — has not yet translated into a perception of reliability or fairness among a vocal segment of its developer base. The infrastructure bottleneck Cherny describes is therefore not just a technical problem but a trust and retention problem: if users experience rate limits and policy reversals during the period when Anthropic is releasing features at its fastest-ever pace, the product momentum risks being undermined by the operational experience surrounding it.

The situation illustrates a structural tension common to AI companies at this stage of growth: the product innovation cycle and the infrastructure maturation cycle operate on different timescales. Feature releases can be shipped in days; the capacity, pricing architecture, and policy frameworks needed to support those features reliably at scale take months to stabilize. Anthropic's internal acknowledgment — working around the clock to improve the stack — suggests awareness of this gap, but the public thread makes clear that users are experiencing the gap in real time. For Anthropic to convert its shipping velocity into durable competitive advantage, the infrastructure and policy experience will need to catch up with the product ambition, particularly for the developer segment that has historically served as the company's most influential distribution channel.

Read original article →